feat: support wandb

[borisdayma]

I've added the barebones for W&B support.

Few notes:

• I've not built the training logic because I feel it's a separate issue. My own structure follows closely run_common_voice with a few tweaks but I see we already created a few utils in this repo.
• I'm not using WANDB_LOG_MODEL and instead manually log all the required files because it would only log the model files while we also want the preprocecssing files (which are not saved with trainer.save_model())
• For doing a parameter search using W&B sweeps, we will just need a yaml like this
• I didn't add wandb into the requirements (should it be dev only) as I'm not sure if this repo will be mainly for training or also for evaluating (I imagine evaluating would be directly through transformers)